By Steven Rosenbush
This week's technology stock selloff, now in its third day, has its roots in a new class of AI agents that emerged a few months ago. These "long-running" agents are still growing in power, too. On Thursday, Anthropic launched an upgraded version of the technology that sparked the selloff.
Companies need to pay attention.
AI agents have the potential to draw value out of almost every sector of the economy, not just software and data. That isn't inevitable, though. Companies can use AI to create value of their own, provided they keep pace with technological changes and their downstream impact on business. For those that fall behind, the dreaded bubble may not be what lies within the AI sector, but what exists beyond.
Investors have been concerned for months that AI could suck the value out of narrowly focused software-as-a-service companies, rendering them mere databases that feed AI agents. It isn't difficult to see how this new generation of agents that run for longer periods could end up stepping between the customer relationships that SaaS companies have cultivated. It could be similar to the way that Apple extracted the value in mobile communications from hardware-focused incumbents that didn't get software.
Those fears were compounded in November by the growing power of Anthropic's Claude Code, which builds software with stunning proficiency. On Jan. 12, the company announced Cowork, a research preview feature in Claude's desktop app that lets users assign multistep tasks to Claude such as organizing files, drafting documents, extracting data from spreadsheets or synthesizing research. Instead of chatting with Claude, users can point it at a folder on their computer and give it an objective.
On Tuesday morning, investors homed in on Anthropic's Friday announcement that it was adding open-source plug-ins in research preview to Cowork that allow companies to customize and build on Claude. The 11 plug-ins have a range of focus, including the legal profession.
It was sufficient to give them a glimpse into the afterworld of economic life as we know it. To avoid irrelevance in the future, companies will need to think carefully about where their competitive advantage lies in a world of shared AI platforms.
Co-what?
Claude Code and Cowork are products that run on multiple Claude models. On Thursday, Anthropic announced a new model, Opus 4.6, that it says is a major step up in professional skills and reasoning compared with Opus 4.5 released in November.
As a result of the recent breakthroughs, investors who spent last year debating the existence of an AI bubble started to take AI's potential more seriously. The Cowork plug-ins were like a match on dry tinder. If AI agents like Cowork can disrupt one research-intensive field such as the law, why not another and another? Suddenly this seemed like a near-term threat, not a speculative problem that's always five years in the future.
I spoke to Scott White, Anthropic's head of product for enterprise, to get a better understanding of how these agents operate and how they stand to transform work, companies and markets.
Going long
Claude Code and Cowork are pioneers of an emerging class of long-running AI agents with implications far beyond any one particular sector such as software development.
Unlike previous iterations that generated responses immediately, developers can now specify the level of effort they want Claude to use for a given request. This allows them to assess the complexity of a requested task and determine how hard it needs to think to solve it, White told me.
That's how the model can engage in "long-horizon" thinking. The fusion of a reasoning model with a capable "harness" allows the model to connect to real systems, execute code and manage workflows, he said. Users can now assign a broad objective to the AI, which will operate autonomously in the background, maintaining context over long horizons. The agent can work for longer periods, but the turnaround time for projects is much faster.
Researchers Alexi Robbins and Jonas Nelle of Cursor said they use Opus 4.5 in conjunction with OpenAI's GPT-5.2. "When these models were stacked together, they maintained coherence over extremely long horizons, allowing for high degrees of complexity," Robbins said.
During the development of Cursor's browser, Robbins said he observed the company's agent intentionally drawing bright boxes and borders around elements on the webpage that weren't part of the design, taking screenshots and debugging on its own. "It invented a tool it wasn't given," Nelle said.
Rethinking business models
White said these new capabilities are already changing the way he and his colleagues work.
In the past, he wrote documents to convince engineers to build features. Now he uses Claude to build working prototypes to demonstrate viability. The near-instantaneous merging of internal data, web research and customer feedback compresses weeks of analysis into minutes of execution.
Anthropic designers now write production code to implement their own interfaces, and product managers are performing complex data-science tasks that previously required specialists, he said.
Work is changing at three levels, White said. Individuals are becoming more capable and productive and companies are tearing down historical workflows from marketing to compliance, reducing turnaround times. "Lastly, businesses are changing how they think about what they're going to look like, what products they want to build," White said. "Building new things is so much more approachable for businesses, introducing new revenue lines."
LegalZoom CEO Jeff Stibel said he approaches AI as an accelerant for his business and so-called Main Street companies. The online legal-services platform has an in-house law firm and an independent network of more than 1,000 attorneys, as well as concierge service agents. It also takes an agnostic approach to client technology.
In Stibel's view, AI makes it easier to start a business, driving clients to specialized services that support them. LegalZoom is moving from a search-and-answer approach to an emphasis on execution and a broader solution that clients trust.
"AI provides the insight, but LegalZoom provides the trusted solution," he said.
A new architecture
The arrival of long-running agents marks a sudden advance in the design of AI systems. To grasp their potential implications, it is helpful to understand something about their architecture. I turned to Alex Salazar, co-founder and CEO of San Francisco-based AI startup Arcade.dev, for insight.
In the old days, before Thanksgiving, he said, a developer might have prompted a chatbot by saying, "You are an accounting agent; here is the enterprise resource planning tool."
The new architecture allows the user to assign the agent a higher-level goal, such as "check depreciation schedules." The model, acting as a planner, "dynamically decides which skills and tools it needs to solve the problem, iterates on the plan, and executes it without rigid preprogramming," Salazar said.
This allows for caching or storing more information in dedicated memory, which makes the agents faster and cheaper because they don't have to reprocess the same data for every single interaction, according to Salazar.
"This shifts the nature of work from creation to editing," he said. He maintains that junior employees must pivot from being the "writer" to the "editor" driving the agent. "The skill set of the future is not syntax or rote creation, but taste and judgment," he said.
And fears about the disappearance of entry-level work notwithstanding, Salazar said junior employees entering the workforce now, who have a native grasp of AI tools, will likely outperform senior employees who insist on doing the grunt work manually.
If long-running agents realize even part of this potential, the impact on the economy will be huge. This week's selloff suggests that investors are beginning to take that prospect much more seriously.
Write to Steven Rosenbush at steven.rosenbush@wsj.com
(END) Dow Jones Newswires
February 05, 2026 13:03 ET (18:03 GMT)
Copyright (c) 2026 Dow Jones & Company, Inc.
Comments